You may have heard of people who add chatbots to controls that do real work. Controls can run Internet searches, run commands to open and read documents and spreadsheets, or even edit Delete entire databaseWhether or not this seems like a good idea depends partly on how bad it is if the chatbot does something destructive, and how destructive you’ve allowed it to be,
That’s why running a single in-house company store is a good test application for such powerful chatbots. Not because AI is likely to do a very good job, but because the disadvantages have been overcome.
Anthropic recently shared an experiment In which he used a chatbot to run his company’s store. A human employee still had to stock the shelves, but they put an AI agent (which they called the cloud) in charge of interacting with customers about the source of the products and then researching the products online. How good did it go? In my opinion, not that good.
Cloud:
- Easily persuaded to offer discounts and free items
- Began stockpiling tungsten cubes upon request and selling them at huge losses
- Invented conversations with employees who didn’t exist
- Claimed to have visited 742 Evergreen Terrace (the fictional address of The Simpsons family)
- He was claimed to be wearing a navy blue blazer and red tie on the site
That was in June. Sometime this year Anthropic convinced Wall Street Journal journalists to try a somewhat updated version of the cloud (which they called Claudius) for an in-house store. His writeup is very funny(original) Herearchived version Here,
In short, Claudius:
- Reassured on several occasions that everything should be given away for free
- Ordered a Playstation 5 (which he gave away for free)
- Ordered a live betta fish (which he gave away for free)
- Told an employee he left a stack of cash for them near the register
- Was highly entertaining. “Profits collapsed. Newsroom morale soared.”
(The betta fish are fine, happily installed in a large tank in the newsroom.)
Why couldn’t chatbots stay connected to reality? Keep in mind that large language models are inherently improving. They will follow their original instructions only as long as following those instructions is most likely to be the next line of the script. Is the script an actual transcript of a model customer service interaction? a science fiction story? Both scenarios are in its Internet training data, and there is no way to tell what the real-world truth is. A newsroom full of talented journalists can easily enable chatbots to change the landscape. I don’t think this problem will go away – it’s quite fundamental to how large language models work.
I want a Claude or Claudius vending machine, but only because it’s weird and entertaining. And of course only if someone else provides the budget.
Bonus material for AI Weirdness supporters: I revisit the Christmas Carol dataset using the small old school language model Char-RNN. Things become scandalous very quickly.
